JP6871352B1 - Learning equipment, learning methods and learning programs - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a value which contributes to the interpretability of a model with an easily observable value while maintaining the accuracy of the model. SOLUTION: When a learning device 10 acquires a plurality of data, inputs the acquired plurality of data to a model as input data, and obtains output data output from the model, it becomes output data and correct answer data. Based on this, calculate the loss of the model. Then, each time the loss is calculated, the learning device 10 repeats the update process of updating the weight of the model according to the loss. Further, the learning device 10 calculates a value that contributes to the interpretability of the model, and ends the update process when the loss and the value that contributes to the interpretability of the model satisfy a predetermined condition. [Selection diagram] Fig. 1

Description

The present invention relates to learning devices, learning methods and learning programs.

Conventionally, a method of extracting a value that contributes to the interpretability of a model has been known. For example, in the case of neural networks, multiple proposals have been made, including the Saliency map, as a method for extracting the input / output relationships of neural networks. It is used for the purpose of showing, and is also used in the actual system. The numerical value of the input / output relationship obtained by this method is calculated by an algorithm using back propagation for each input sample with respect to the trained model of the neural network.

In addition, even in cases other than neural networks, Importance indicates the contribution obtained by LIMITE and SHAP that can be used for any model, and the importance of input obtained by a method using a decision tree such as Gradient Boosting Tree. Score is also used as an interpretation of the model. The values that contribute to the interpretability of these models are hereinafter referred to as attribution.

Smilkov Daniel, et al. “Smoothgrad: removing noise by adding noise.” ArXiv preprint 1706.03825 (2017). Simonyan Karen, Andrea Vedaldi, and Andrew Zisserman. “Deep inside convolutional networks: Visualising image classification models and saliency maps.” ArXiv preprint arXiv: 1312.6034 (2014). Binder Alexander, et al. “Layer-wise relevance propagation for deep neural network architectures.” Information Science and Applications (ICISA) 2016. Springer, Singapore, 2016. 913-922. Ribeiro Marco Tulio, Sameer Singh, and Carlos Guestrin. "Why should i trust you ?: Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. ACM, 2016. Strumbelj Erik, and Igor Kononenko. "Explaining prediction models and individual predictions with feature contributions." Knowledge and information systems 41.3 (2013): 647-665.

However, for a machine learning model in which the condition of the number of learnings is specified and sequential learning is performed, it is difficult to obtain a value that contributes to the interpretability of the model with an easily observable value by the conventional learning method. was there. For example, since the value obtained as an attribution depends on the learning process of the model, the attribution obtained from the model after a certain number of learnings shows the relationship between input and output in an interpretable form (). (Hereafter, it is called that attribution converges), and there are cases where noise is included and it is difficult to understand, so it was difficult to stabilize.

This is because the end criterion for model learning is fixed number of trainings, training is stopped based on whether the accuracy represented by Early Stopping improves, or the accuracy is constant like hyperparameter search. This is because many of them utilize exceeding the above, and do not guarantee that attribution can be obtained without noise.

In order to solve the above-mentioned problems and achieve the object, the learning device of the present invention inputs an acquisition unit for acquiring a plurality of data and a plurality of data acquired by the acquisition unit into a model as input data. When the output data output from the model is obtained, the loss is calculated by the first calculation unit that calculates the loss of the model and the first calculation unit based on the output data and the correct answer data. Each time, an update unit that repeatedly updates the weight of the model according to the loss, a second calculation unit that calculates a value that contributes to the interpretability of the model, and the first calculation unit It is characterized by having an update end unit that terminates the update process when the calculated loss and the value calculated by the second calculation unit satisfy a predetermined condition.

Further, the learning method of the present invention is a learning method executed by a learning device, in which an acquisition step of acquiring a plurality of data and a plurality of data acquired by the acquisition step are input to a model as input data. When the output data output from the model is obtained, the loss is calculated by the first calculation step of calculating the loss of the model and the first calculation step based on the output data and the correct answer data. By the update process of repeating the update process of updating the weight of the model according to the loss, the second calculation process of calculating the value contributing to the interpretability of the model, and the first calculation process of each time. When the calculated loss and the value calculated by the second calculation step satisfy a predetermined condition, the update end step of ending the update process is included.

Further, in the learning program of the present invention, when a acquisition step for acquiring a plurality of data and a plurality of data acquired by the acquisition step are input to a model as input data and output data output from the model is obtained. In addition, the first calculation step for calculating the loss of the model based on the output data and the correct answer data, and each time the loss is calculated by the first calculation step, the weight of the model is weighted according to the loss. By the update step of repeating the update process, the second calculation step of calculating the value contributing to the interpretability of the model, the loss calculated by the first calculation step, and the second calculation step. When the calculated value satisfies a predetermined condition, the computer is made to execute the update end step for ending the update process.

According to the present invention, it is possible to obtain a value that contributes to the interpretability of the model with an easily observable value while maintaining the accuracy of the model.

FIG. 1 is a block diagram showing a configuration example of the learning device according to the first embodiment. FIG. 2 is a diagram illustrating an outline of a learning process executed by the learning device. FIG. 3 is a flowchart showing an example of the flow of learning processing in the learning device according to the first embodiment. FIG. 4 is a block diagram showing a configuration example of the learning device according to the second embodiment. FIG. 5 is a diagram illustrating an outline of an abnormality prediction process and an attribution extraction process executed by the learning device. FIG. 6 is a diagram illustrating an outline of an image classification process and an attribution extraction process executed by the learning device. FIG. 7 is a flowchart showing an example of the flow of attribution extraction processing in the learning device according to the second embodiment. FIG. 8 is a diagram showing a computer that executes a program.

Hereinafter, the learning device, the learning method, and the embodiment of the learning program according to the present application will be described in detail with reference to the drawings. The learning device, learning method, and learning program according to the present application are not limited by this embodiment.

[First Embodiment]
In the following embodiments, the configuration of the learning device 10 and the processing flow of the learning device 10 according to the first embodiment will be described in order, and finally, the effects of the first embodiment will be described.

[Configuration of learning device]
First, the configuration of the learning device 10 will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration example of the learning device according to the first embodiment. The learning device 10 uses the learning data prepared in advance to perform a learning process of repeating the process of updating the weight of the model. In the learning device 10, not only the accuracy of the model but also the attribution value is taken into consideration in the learning end condition in order to guarantee that the attribution noise is reduced in the learning process. For example, the learning device 10 applies a scale for measuring the sparseness of attribution (for example, the L1 norm of the attribution score or the GINI coefficient of the attribution score) to the learning end condition, and the accuracy is reduced to a certain value or less. Learning can be terminated when the degree of sparseness exceeds a certain value.

As shown in FIG. 1, the learning device 10 includes a communication processing unit 11, a control unit 12, and a storage unit 13. The processing of each part of the learning device 10 will be described below.

The communication processing unit 11 controls communication regarding various information exchanged with the connected device. Further, the storage unit 13 stores data and programs required for various processes by the control unit 12, and has a data storage unit 13a and a learned model storage unit 13b. For example, the storage unit 13 is a storage device such as a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory).

The data storage unit 13a stores data acquired by the acquisition unit 12a, which will be described later. For example, the data storage unit 13a stores a learning data set to which a correct answer label is given in advance. The type of data may be any data as long as it is data consisting of a plurality of real values. For example, data of a sensor provided in a target device such as a factory, a plant, a building, or a data center ( For example, it may be data such as temperature, pressure, sound, vibration, etc.), or it may be image data data.

The learned model storage unit 13b stores the learned model learned by the learning process described later. For example, the trained model storage unit 13b stores the prediction model of the neural network for predicting the abnormality of the monitored equipment as the trained model.

The control unit 12 has an internal memory for storing a program that defines various processing procedures and the like and required data, and executes various processing by these. For example, the control unit 12 has an acquisition unit 12a, a first calculation unit 12b, an update unit 12c, a second calculation unit 12d, and an update end unit 12e. Here, the control unit 12 is, for example, an electronic circuit such as a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphical Processing Unit), an ASIC (Application Specific Integrated Circuit), or an FPGA (Field Programmable Gate Array). It is an integrated circuit such as.

The acquisition unit 12a acquires a plurality of data. For example, the acquisition unit 12a reads and acquires the data set stored in the data storage unit 13a. Here, the data acquired by the sensor is, for example, various data such as temperature, pressure, sound, and vibration of the equipment to be monitored, the equipment in the plant, and the reactor. Further, the data acquired by the acquisition unit 12a is not limited to the data acquired by the sensor, and may be, for example, image data, human-input numerical data, or the like. The acquisition unit 12a may acquire data in real time. For example, the acquisition unit 12a may periodically (for example, every minute) acquire the numerical data of the multivariate time series from the sensor installed in the monitored equipment such as a factory or a plant.

The first calculation unit 12b inputs a plurality of data acquired by the acquisition unit 12a into the model as input data, and when the output data output from the model is obtained, the first calculation unit 12b is based on the output data and the correct answer data. Calculate model loss. For example, the first calculation unit 12b calculates the loss of the model using a predetermined loss function. The loss calculation method is not limited, and any method may be used.

Each time the loss is calculated by the first calculation unit 12b, the update unit 12c repeats the update process of updating the weight of the model according to the loss. The update unit 12c updates the weight (parameter) according to the magnitude of the loss. The update method is not limited, and any method may be used.

The second calculation unit 12d calculates a value that contributes to the interpretability of the model. For example, the second calculation unit 12d calculates the attribution, which is the contribution of each element of the input data to the output data, based on the input data and the output data.

Here, a specific example of calculating attribution will be described. For example, the second calculation unit 12d uses the partial differential value or its approximate value for each input value of the output value in the trained model that calculates the output value from the input value, and attributs each sensor at each time. Calculate the input. As an example, the second calculation unit 12d uses the Saliency Map to calculate the attribution for each sensor at each time. Saliency Map is a technique used in the image classification of a neural network, and is a technique for extracting a partial differential value for each input of the output of the neural network as an attribution that contributes to the output. Attribution may be calculated by a method other than Saliency Map.

Further, the value that contributes to the interpretability of the model calculated by the second calculation unit 12d is not limited to attribution, and may be, for example, a value that represents the sparsity of the weight of the model.

The update end unit 12e ends the update process when the loss calculated by the first calculation unit 12b and the value calculated by the second calculation unit 12d satisfy a predetermined condition. For example, in the update end unit 12e, the loss calculated by the first calculation unit 12b is equal to or less than the preset threshold value, and the value calculated by the second calculation unit 12d is equal to or less than the preset threshold value. If is, the update process may be terminated. More specifically, the update end unit 12e ends the update process when the loss is equal to or less than a predetermined threshold value and the L1 norm of attribution is equal to or less than a preset threshold value.

Further, in the update end unit 12e, in the update end unit, the loss calculated by the first calculation unit 12b continues to be larger than the loss calculated last time for a predetermined number of times in a row, and the second calculation unit 12d If the value calculated by is larger than the previously calculated value for a predetermined number of consecutive times, the update process may be terminated. More specifically, in the update end portion 12e, the loss becomes larger than the previously calculated loss five times in a row, and the L1 norm of the attribution is the L1 norm of the previously calculated attribution. If the value becomes larger than 5 times in a row, the update process may be terminated.

Here, the outline of the learning process executed by the learning device 10 will be described with reference to FIG. FIG. 2 is a diagram illustrating an outline of a learning process executed by the learning device. As illustrated in FIG. 2, the learning device 10 learns the model by repeatedly performing Phase 1 for updating the weight and Phase 2 for calculating the attribution. Further, the learning device 10 determines whether to end the learning based on the calculated loss and attribution values.

In Phase 1, the learning device 10 inputs the learning data to the model, acquires the output data output from the model, calculates the loss based on the output data and the correct label, and determines the loss according to the magnitude of the loss. Update the weight.

Subsequently, in Phase 2, the learning device 10 inputs the verification data into the model, acquires the output data output from the model, and calculates the attribution based on the input data and the output data. Further, the learning device 10 calculates the loss based on the output data and the correct answer label. Here, the verification data may be the same data as the training data input to the model in Phase 1, or may be different data.

Then, the learning device 10 determines whether or not to end the learning based on the calculated loss and attribution values. For example, the learning device 10 ends the update process when the loss is equal to or less than a predetermined threshold value and the L1 norm of attribution is equal to or less than a preset threshold value.

When attribution is used as a value that contributes to the interpretability of the model, the learning device 10 calculates the L1 norm of the attribution by, for example, the following equation (1). In the following calculation formula, "x _ij " is a value of the input data sample i and the feature j. Further, in the following calculation formula, "A" is a function for calculating attribution from features and models, and "M" is a model.

Further, the learning device 10 may terminate the update process when the loss is equal to or less than a predetermined threshold value and the L1 norm of the model weight is equal to or less than a preset threshold value. For example, when the learning device 10 uses the model weight L1 norm as a value other than attribution as a value that contributes to the interpretability of the model, for example, the model weight L1 is calculated by the following equation (2). Calculate the norm. In the following formula, "x _ijk " means the weight from node j to node k in the i-layer of the model.

As a result, when the learning device 10 determines that the learning is completed, the learning device 10 outputs the learned model or stores the learned model in the learned model storage unit 13b. Further, when it is determined that the learning is completed, the learning device 10 returns to Phase 1 and performs a process of updating the weight. That is, the learning device 10 learns the model by repeatedly performing Phase 1 for updating the weight and Phase 2 for calculating the attribution until it is determined that the learning is completed.

As described above, in the learning device 10, not only the accuracy of the model but also the attribution value is introduced into the learning end condition in order to guarantee that the attribution noise is reduced in the learning. For example, in the learning device 10, a scale for measuring the sparsity of attribution is applied to the learning end condition, and learning is ended when the accuracy is equal to or less than a certain value and the sparsity degree is also equal to or more than a certain value. be able to.

Further, in the learning device 10, since the value of the attribution is directly included in the learning end condition, it is possible to consider the convergence of the attribution, which has not been guaranteed by the learning using only the accuracy as the end condition. The stability of the resulting attribution score can be increased.

In addition, since the learning curve has the characteristic of repeating the stagnation and descent of the loss depending on the data, there is a problem that the conventional Early Stopping, which only looks at the accuracy, pays attention to the learning before the loss actually converges. On the other hand, it is known that there is a close relationship between the end of learning and the convergence of attribution, and in the learning device 10, by including the convergence of attribution as the end condition, the above-mentioned learning curve is stagnant. Even at times, if the attribution does not converge, it is possible to obtain a judgment that does not stop learning.

The model of this embodiment may be a model other than the neural network. For example, in addition to the neural network, there are some models such as Gradient Boosting that sequentially perform learning by using the gradient descent method or the like, and this embodiment can also be used for these models. In the learning device 10, LIME and SHAP exist as a method that can take out the input / output relationship for any model in general. By calculating this value at the time of learning, it is possible to realize a mechanism that stops learning when it becomes sparse, similar to attribution (formula). In addition, a method such as Gradient Boosting Decision Tree can calculate the importance score of each feature. By using this score in the same way as the weight, it is possible to realize a mechanism that stops learning when it becomes sparse like the weight (formula).

[Processing procedure of learning device]
Next, an example of the processing procedure by the learning device 10 according to the first embodiment will be described with reference to FIG. FIG. 3 is a flowchart showing an example of the flow of learning processing in the learning device according to the first embodiment. In the example of FIG. 3, a case where attribution is used as a value that contributes to the interpretability of the model will be described as an example.

As illustrated in FIG. 3, the acquisition unit 12a of the learning device 10 acquires data. For example, the acquisition unit 12a reads and acquires the data set stored in the data storage unit 13a (step S101). Then, the first calculation unit 12b inputs the data acquired by the acquisition unit 12a into the model (step S102), and calculates the loss of the model based on the output data and the correct answer data (step S103).

Then, the update unit 12c updates the weight of the model according to the loss loss by the first calculation unit 12b (step S104). Subsequently, the second calculation unit 12d calculates the attribution using the input data and the output data (step S105). For example, when the second calculation unit 12d inputs a plurality of sensor data as input data into a prediction model for predicting the state of the monitored equipment and obtains output data output from the prediction model. , Calculate attribution per sensor based on input and output data.

Then, the update end unit 12e determines whether or not the loss calculated by the first calculation unit 12b and the attribution calculated by the second calculation unit 12d satisfy a predetermined condition (step S106). .. For example, the update end unit 12e determines whether the update end unit 12e determines whether or not the loss is equal to or less than a predetermined threshold value and the attribution L1 norm is equal to or less than a preset threshold value.

As a result, when the update end unit 12e determines that the loss and the attribution do not satisfy the predetermined conditions (step S106 is denied), the learning device 10 returns to the process of step S101 and returns to the process of the loss and attribution. The process of steps S101 to S106 is repeated until the above conditions are satisfied.

When the update end unit 12e determines that the loss and the attribution satisfy a predetermined condition (affirmation in step S106), the trained model is stored in the trained model storage unit 13b (step S107).

[Effect of the first embodiment]
The learning device 10 according to the first embodiment acquires a plurality of data, inputs the acquired plurality of data to the model as input data, and when the output data output from the model is obtained, the output data and the output data are obtained. Calculate the model loss based on the correct answer data. Then, each time the loss is calculated, the learning device 10 repeats the update process of updating the weight of the model according to the loss. Further, the learning device 10 calculates a value that contributes to the interpretability of the model, and ends the update process when the loss and the value that contributes to the interpretability of the model satisfy a predetermined condition. Therefore, in the learning device 10, it is possible to obtain a value that contributes to the interpretability of the model with an easily observable value while maintaining the accuracy of the model.

That is, in the learning device 10 according to the first embodiment, unlike the learning end condition conventionally used, by adding the attribution value to the learning end condition, for example, the attribution of the trained model It is possible to reduce the noise of. A state in which attribution noise is reduced is a sparse and smooth state that is easy for the observer to observe. Further, in the learning device 10 according to the first embodiment, by adding an attribution value to the learning end condition as compared with the conventionally used learning end condition, for example, based on the accuracy such as Early Stopping. In the method of stopping learning, it is possible to take measures that do not stop learning even when learning is stagnant.

[Second Embodiment]
In the first embodiment described above, the learning device for learning the model has been described, but in the second embodiment, the learning device for extracting attributions using the learned model obtained by the learning process will be described. To do. In the second embodiment below, the configuration of the learning device 10A and the processing flow of the learning device 10A according to the second embodiment will be described in order, and finally the effect of the first embodiment will be described. The description of the same configuration and processing as in the first embodiment will be omitted.

[Configuration of learning device]
First, the configuration of the learning device 10A will be described with reference to FIG. FIG. 4 is a block diagram showing a configuration example of the learning device according to the second embodiment. The learning device 10A collects a plurality of data acquired by sensors installed in the monitored equipment such as a factory or a plant, and inputs the collected data as an input to predict an abnormality in the monitored equipment. The estimated value of the specific sensor of the monitored equipment is output using the trained model of. Further, the learning device 10A may calculate the degree of abnormality from the estimated value output in this way.

For example, the degree of anomaly should be defined as the error between the estimated value of the sensor output by the model and the specific value specified in advance when a regression model with the value of a specific sensor as the objective variable is learned. Is possible. Alternatively, when the model is learned by treating the presence or absence of an abnormality as a classification problem, the ratio of the time zones classified as abnormal within the specified time can be used. Further, the learning device 10A calculates the attribution which is the contribution to the output value of each sensor by using the data of each sensor input to the trained model and the output data output from the trained model. .. Here, the attribution indicates how much each input contributed to the output, and the larger the absolute value of the attribution, the higher the influence of the input on the output. ..

The learning device 10A has a communication processing unit 11, a control unit 12, and a storage unit 13. The control unit 12 includes an acquisition unit 12a, a first calculation unit 12b, an update unit 12c, a second calculation unit 12d, an update end unit 12e, an extraction unit 12f, a prediction unit 12g, and a visualization unit 12h. Here, the learning device 10A is different from the learning device 10 in that it further includes an extraction unit 12f, a prediction unit 12g, and a visualization unit 12h. Regarding the acquisition unit 12a, the first calculation unit 12b, the update unit 12c, the second calculation unit 12d, and the update end unit 12e, the acquisition unit 12a and the first unit of the learning device 10 described in the first embodiment. Since the same processing as that of the calculation unit 12b, the update unit 12c, the second calculation unit 12d, and the update end unit 12e is performed, the description thereof will be omitted.

The extraction unit 12f inputs input data to the trained model in which the update process is repeatedly performed by the update unit 12c until the update process is completed by the update end unit 12e, and obtains output data output from the trained model. If so, the values that contribute to the interpretability of the model are extracted. For example, the extraction unit 12f reads the trained model from the trained model storage unit 13b, inputs the data to be processed into the trained model, and extracts the attribution for each data.

For example, in the trained model that calculates the output value from the input value, the extraction unit 12f calculates the attribution for each sensor at each time using the partial differential value or its approximate value for each input value of the output value. To do. As an example, the extraction unit 12f uses the Saliency Map to calculate the attribution for each sensor at each time.

The prediction unit 12g inputs a plurality of data and outputs a predetermined output value by using, for example, a trained model for predicting the state of the monitored equipment. For example, the prediction unit 12g calculates the degree of abnormality of the monitored equipment using the process data and the trained model (discrimination function or regression function), and predicts whether or not the abnormality will occur after a predetermined fixed time. To do.

The visualization unit 12h visualizes the attribution extracted by the extraction unit 12f and the degree of abnormality calculated by the prediction unit 12g. For example, the visualization unit 12h displays a graph showing the transition of attribution of each sensor data, and displays the calculated abnormality degree as a chart screen.

Here, the outline of the abnormality prediction process and the attribution extraction process executed by the learning device 10A will be described with reference to FIG. FIG. 5 is a diagram illustrating an outline of an abnormality prediction process and an attribution extraction process executed by the learning device.

FIG. 5 shows that a sensor, a device for collecting operating signals, and the like are attached to a reactor, an apparatus, or the like in a plant, and data is collected at regular intervals. Then, FIG. 6 illustrates the transition of the process data collected from each of the sensors A to E, and as described in the first embodiment, the trained model is obtained by learning the model. Generate. Then, the prediction unit 12g predicts the abnormality after a certain period of time using the trained model. Then, the visualization unit 12h outputs the calculated time-series data of the degree of abnormality as a chart screen.

Further, the extraction unit 12f extracts an attribution to a predetermined output value for each sensor at each time using the process data input to the trained model and the output value from the trained model. Then, the visualization unit 12h displays a graph showing the transition of the importance of the process data of each sensor with respect to the prediction.

Further, the learning device 10A is not applied only to the abnormality prediction processing, and for example, image data may be collected and applied to the image classification processing. Here, the outline of the image classification process and the attribution extraction process executed by the learning device 10A will be described with reference to FIG. FIG. 6 is a diagram illustrating an outline of an image classification process and an attribution extraction process executed by the learning device.

In FIG. 6, image data is collected, and the collected image data is used as input data to generate a trained model by training the model as described in the first embodiment. Then, the prediction unit 12g classifies the images included in the image data using the trained model. For example, in the example of FIG. 6, the prediction unit 12g determines whether the image included in the image data is a car image or an airplane image, and outputs the determination result.

Further, the extraction unit 12f extracts the attribution for each pixel in each image by using the image data input to the trained model and the classification result output from the trained model. Then, the visualization unit 12h displays an image showing the attribution for each pixel in each image. In this image, the attribution is expressed by shading. The larger the attribution, the darker the predetermined color, and the smaller the attribution, the lighter the predetermined color.

[Processing procedure of learning device]
Next, an example of the processing procedure by the learning device 10A according to the second embodiment will be described with reference to FIG. 7. FIG. 7 is a flowchart showing an example of the flow of attribution extraction processing in the learning device according to the second embodiment.

As illustrated in FIG. 7, when the extraction unit 12f of the learning device 10 acquires the data (affirmation in step S201), the input data is input to the trained model (step S202), and the output output from the trained model is output. When the data is obtained, the attribution is calculated using the input data and the output data (step S203).

Then, the visualization unit 12h displays a graph that visualizes the attribution (step S204). For example, the visualization unit 12h displays a graph showing the transition of attribution of each sensor data.

As described above, the learning device 10A according to the second embodiment inputs the input data to the trained model learned by the learning process described in the first embodiment, and the output data output from the trained model. When obtained, the attribution to the output data of each element of the input data is extracted based on the input data and the output data. Therefore, the learning device 10A can extract the attribution with less noise.

[System configuration, etc.]
Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device is realized by a CPU or GPU and a program that is analyzed and executed by the CPU or GPU, or as hardware by wired logic. It can be realized.

Further, among the processes described in the present embodiment, all or part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed can be performed. All or part of it can be done automatically by a known method. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[program]
It is also possible to create a program in which the processing executed by the information processing apparatus described in the above embodiment is described in a language that can be executed by a computer. For example, it is possible to create a program in which the processes executed by the learning devices 10 and 10A according to the embodiment are described in a language that can be executed by a computer. In this case, the same effect as that of the above embodiment can be obtained by executing the program by the computer. Further, the same processing as that of the above-described embodiment may be realized by recording the program on a computer-readable recording medium, reading the program recorded on the recording medium into the computer, and executing the program.

FIG. 8 is a diagram showing a computer that executes a program. As illustrated in FIG. 8, the computer 1000 has, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. However, each of these parts is connected by a bus 1080.

The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012, as illustrated in FIG. The ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090, as illustrated in FIG. The disk drive interface 1040 is connected to the disk drive 1100 as illustrated in FIG. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120, as illustrated in FIG. The video adapter 1060 is connected, for example, to a display 1130, as illustrated in FIG.

Here, as illustrated in FIG. 8, the hard disk drive 1090 stores, for example, the OS 1091, the application program 1092, the program module 1093, and the program data 1094. That is, the above-mentioned program is stored in, for example, the hard disk drive 1090 as a program module in which instructions executed by the computer 1000 are described.

Further, the various data described in the above embodiment are stored as program data in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes various processing procedures.

The program module 1093 and program data 1094 related to the program are not limited to the case where they are stored in the hard disk drive 1090, and may be stored in, for example, a removable storage medium and read by the CPU 1020 via a disk drive or the like. .. Alternatively, the program module 1093 and the program data 1094 related to the program are stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.), and are stored via the network interface 1070. It may be read by the CPU 1020.

The above-described embodiments and modifications thereof are included in the inventions described in the claims and the equivalent scope thereof, as are included in the technology disclosed in the present application.

10, 10A Learning device 11 Communication processing unit 12 Control unit 12a Acquisition unit 12b First calculation unit 12c Update unit 12d Second calculation unit 12e Update end unit 12f Extraction unit 12g Prediction unit 12h Visualization unit 13 Storage unit 13a Data storage unit 13b Learned model storage

Claims (7)

An acquisition unit that acquires multiple data, and
When a plurality of data acquired by the acquisition unit are input to the model as input data and output data output from the model is obtained, the loss of the model is calculated based on the output data and the correct answer data. The first calculation unit and
Each time a loss is calculated by the first calculation unit, an update unit that repeatedly performs an update process for updating the weight of the model according to the loss, and an update unit.
A second calculation unit that calculates values that contribute to the interpretability of the model,
When the loss calculated by the first calculation unit and the value calculated by the second calculation unit satisfy a predetermined condition, the update end unit that terminates the update process and the update end unit.
A learning device characterized by having. The learning according to claim 1, wherein the second calculation unit calculates an attribution, which is a contribution of each element of the input data to the output data, based on the input data and the output data. apparatus. When the loss calculated by the first calculation unit is equal to or less than a predetermined threshold value and the value calculated by the second calculation unit is equal to or less than a predetermined threshold value, the update end unit is used. The learning device according to claim 1, wherein the update process is terminated. In the update end unit, the loss calculated by the first calculation unit continues to be larger than the loss calculated last time for a predetermined number of times in a row, and the value calculated by the second calculation unit is the previous value. The learning device according to claim 1, wherein when the value becomes larger than the calculated value continuously for a predetermined number of times, the update process is terminated. When the input data is input to the trained model in which the update process is repeatedly performed by the update unit until the update process is completed by the update end unit, and the output data output from the trained model is obtained, the above-mentioned The learning device according to claim 1, further comprising an extraction unit for extracting values that contribute to the interpretability of the model. A learning method performed by a learning device,
The acquisition process to acquire multiple data and
When a plurality of data acquired in the acquisition process are input to the model as input data and output data output from the model is obtained, the loss of the model is calculated based on the output data and the correct answer data. The first calculation process and
Each time a loss is calculated by the first calculation step, an update process of repeatedly updating the weight of the model according to the loss, and an update step of repeating the update process.
A second calculation step for calculating values that contribute to the interpretability of the model,
When the loss calculated by the first calculation step and the value calculated by the second calculation step satisfy a predetermined condition, the update end step of ending the update process and the update end step of ending the update process.
A learning method characterized by including. The acquisition step to acquire multiple data and
When a plurality of data acquired by the acquisition step is input to the model as input data and output data output from the model is obtained, the loss of the model is calculated based on the output data and the correct answer data. The first calculation step and
Each time a loss is calculated by the first calculation step, an update step of repeating an update process for updating the weight of the model according to the loss, and an update step.
A second calculation step to calculate the values that contribute to the interpretability of the model,
When the loss calculated by the first calculation step and the value calculated by the second calculation step satisfy a predetermined condition, the update end step for ending the update process and the update end step
A learning program characterized by having a computer execute.

Images